Table 2.

Different criteria for filtering clusters for function prediction

	(Filter 1)	(Filter 1 & Filter 2)	(Filter 1 & Filter 3)	(Filter 1 & Filter 4)	(Filter 1 & Filter 5)
# of groups	196	74	53	185	11
# of terms	345	159	102	338	16
# of genes	3213	711	409	2895	320
Precision	67.91%	62.52%	60.52%	67.73%	64.70%
Recall	22.98%	26.16%	19.78%	23.80%	11.21%

In order to push the values for precision and recall towards the precision ceiling, we strived for filter criteria for selecting appropriate gene groups a-priori. To achieve this goal, we defined the following filter criteria for our 1,000 'phenoclusters':

Filter 1: Removes groups with less than 3 genes, no GO-terms associated to at least 50% of genes

Filter 2: Removes groups with a GO-similarity score < 0.4

Filter 3: Removes groups with a PPi-connectedness < 33%.

Filter 4: removes all non-single species clusters.

Filter 5: removes all single-species clusters